fs/bcachefs/fs-common.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "acl.h"
   5 #include "btree_update.h"
   6 #include "dirent.h"
   7 #include "fs-common.h"
   8 #include "inode.h"
   9 #include "subvolume.h"
  10 #include "xattr.h"
  11
  12 #include <linux/posix_acl.h>
  13
  14 static inline int is_subdir_for_nlink(struct bch_inode_unpacked *inode)
  15 {
  16         return S_ISDIR(inode->bi_mode) && !inode->bi_subvol;
  17 }
  18
  19 int bch2_create_trans(struct btree_trans *trans,
  20                       subvol_inum dir,
  21                       struct bch_inode_unpacked *dir_u,
  22                       struct bch_inode_unpacked *new_inode,
  23                       const struct qstr *name,
  24                       uid_t uid, gid_t gid, umode_t mode, dev_t rdev,
  25                       struct posix_acl *default_acl,
  26                       struct posix_acl *acl,
  27                       subvol_inum snapshot_src,
  28                       unsigned flags)
  29 {
  30         struct bch_fs *c = trans->c;
  31         struct btree_iter dir_iter = { NULL };
  32         struct btree_iter inode_iter = { NULL };
  33         subvol_inum new_inum = dir;
  34         u64 now = bch2_current_time(c);
  35         u64 cpu = raw_smp_processor_id();
  36         u64 dir_target;
  37         u32 snapshot;
  38         unsigned dir_type = mode_to_type(mode);
  39         int ret;
  40
  41         ret = bch2_subvolume_get_snapshot(trans, dir.subvol, &snapshot);
  42         if (ret)
  43                 goto err;
  44
  45         ret = bch2_inode_peek(trans, &dir_iter, dir_u, dir,
  46                               BTREE_ITER_intent|BTREE_ITER_with_updates);
  47         if (ret)
  48                 goto err;
  49
  50         if (!(flags & BCH_CREATE_SNAPSHOT)) {
  51                 /* Normal create path - allocate a new inode: */
  52                 bch2_inode_init_late(new_inode, now, uid, gid, mode, rdev, dir_u);
  53
  54                 if (flags & BCH_CREATE_TMPFILE)
  55                         new_inode->bi_flags |= BCH_INODE_unlinked;
  56
  57                 ret = bch2_inode_create(trans, &inode_iter, new_inode, snapshot, cpu);
  58                 if (ret)
  59                         goto err;
  60
  61                 snapshot_src = (subvol_inum) { 0 };
  62         } else {
  63                 /*
  64                  * Creating a snapshot - we're not allocating a new inode, but
  65                  * we do have to lookup the root inode of the subvolume we're
  66                  * snapshotting and update it (in the new snapshot):
  67                  */
  68
  69                 if (!snapshot_src.inum) {
  70                         /* Inode wasn't specified, just snapshot: */
  71                         struct bch_subvolume s;
  72
  73                         ret = bch2_subvolume_get(trans, snapshot_src.subvol, true,
  74                                                  BTREE_ITER_cached, &s);
  75                         if (ret)
  76                                 goto err;
  77
  78                         snapshot_src.inum = le64_to_cpu(s.inode);
  79                 }
  80
  81                 ret = bch2_inode_peek(trans, &inode_iter, new_inode, snapshot_src,
  82                                       BTREE_ITER_intent);
  83                 if (ret)
  84                         goto err;
  85
  86                 if (new_inode->bi_subvol != snapshot_src.subvol) {
  87                         /* Not a subvolume root: */
  88                         ret = -EINVAL;
  89                         goto err;
  90                 }
  91
  92                 /*
  93                  * If we're not root, we have to own the subvolume being
  94                  * snapshotted:
  95                  */
  96                 if (uid && new_inode->bi_uid != uid) {
  97                         ret = -EPERM;
  98                         goto err;
  99                 }
 100
 101                 flags |= BCH_CREATE_SUBVOL;
 102         }
 103
 104         new_inum.inum   = new_inode->bi_inum;
 105         dir_target      = new_inode->bi_inum;
 106
 107         if (flags & BCH_CREATE_SUBVOL) {
 108                 u32 new_subvol, dir_snapshot;
 109
 110                 ret = bch2_subvolume_create(trans, new_inode->bi_inum,
 111                                             dir.subvol,
 112                                             snapshot_src.subvol,
 113                                             &new_subvol, &snapshot,
 114                                             (flags & BCH_CREATE_SNAPSHOT_RO) != 0);
 115                 if (ret)
 116                         goto err;
 117
 118                 new_inode->bi_parent_subvol     = dir.subvol;
 119                 new_inode->bi_subvol            = new_subvol;
 120                 new_inum.subvol                 = new_subvol;
 121                 dir_target                      = new_subvol;
 122                 dir_type                        = DT_SUBVOL;
 123
 124                 ret = bch2_subvolume_get_snapshot(trans, dir.subvol, &dir_snapshot);
 125                 if (ret)
 126                         goto err;
 127
 128                 bch2_btree_iter_set_snapshot(&dir_iter, dir_snapshot);
 129                 ret = bch2_btree_iter_traverse(&dir_iter);
 130                 if (ret)
 131                         goto err;
 132         }
 133
 134         if (!(flags & BCH_CREATE_SNAPSHOT)) {
 135                 if (default_acl) {
 136                         ret = bch2_set_acl_trans(trans, new_inum, new_inode,
 137                                                  default_acl, ACL_TYPE_DEFAULT);
 138                         if (ret)
 139                                 goto err;
 140                 }
 141
 142                 if (acl) {
 143                         ret = bch2_set_acl_trans(trans, new_inum, new_inode,
 144                                                  acl, ACL_TYPE_ACCESS);
 145                         if (ret)
 146                                 goto err;
 147                 }
 148         }
 149
 150         if (!(flags & BCH_CREATE_TMPFILE)) {
 151                 struct bch_hash_info dir_hash = bch2_hash_info_init(c, dir_u);
 152                 u64 dir_offset;
 153
 154                 if (is_subdir_for_nlink(new_inode))
 155                         dir_u->bi_nlink++;
 156                 dir_u->bi_mtime = dir_u->bi_ctime = now;
 157
 158                 ret = bch2_inode_write(trans, &dir_iter, dir_u);
 159                 if (ret)
 160                         goto err;
 161
 162                 ret = bch2_dirent_create(trans, dir, &dir_hash,
 163                                          dir_type,
 164                                          name,
 165                                          dir_target,
 166                                          &dir_offset,
 167                                          STR_HASH_must_create|BTREE_ITER_with_updates);
 168                 if (ret)
 169                         goto err;
 170
 171                 new_inode->bi_dir               = dir_u->bi_inum;
 172                 new_inode->bi_dir_offset        = dir_offset;
 173         }
 174
 175         inode_iter.flags &= ~BTREE_ITER_all_snapshots;
 176         bch2_btree_iter_set_snapshot(&inode_iter, snapshot);
 177
 178         ret   = bch2_btree_iter_traverse(&inode_iter) ?:
 179                 bch2_inode_write(trans, &inode_iter, new_inode);
 180 err:
 181         bch2_trans_iter_exit(trans, &inode_iter);
 182         bch2_trans_iter_exit(trans, &dir_iter);
 183         return ret;
 184 }
 185
 186 int bch2_link_trans(struct btree_trans *trans,
 187                     subvol_inum dir,  struct bch_inode_unpacked *dir_u,
 188                     subvol_inum inum, struct bch_inode_unpacked *inode_u,
 189                     const struct qstr *name)
 190 {
 191         struct bch_fs *c = trans->c;
 192         struct btree_iter dir_iter = { NULL };
 193         struct btree_iter inode_iter = { NULL };
 194         struct bch_hash_info dir_hash;
 195         u64 now = bch2_current_time(c);
 196         u64 dir_offset = 0;
 197         int ret;
 198
 199         if (dir.subvol != inum.subvol)
 200                 return -EXDEV;
 201
 202         ret = bch2_inode_peek(trans, &inode_iter, inode_u, inum, BTREE_ITER_intent);
 203         if (ret)
 204                 return ret;
 205
 206         inode_u->bi_ctime = now;
 207         ret = bch2_inode_nlink_inc(inode_u);
 208         if (ret)
 209                 goto err;
 210
 211         ret = bch2_inode_peek(trans, &dir_iter, dir_u, dir, BTREE_ITER_intent);
 212         if (ret)
 213                 goto err;
 214
 215         if (bch2_reinherit_attrs(inode_u, dir_u)) {
 216                 ret = -EXDEV;
 217                 goto err;
 218         }
 219
 220         dir_u->bi_mtime = dir_u->bi_ctime = now;
 221
 222         dir_hash = bch2_hash_info_init(c, dir_u);
 223
 224         ret = bch2_dirent_create(trans, dir, &dir_hash,
 225                                  mode_to_type(inode_u->bi_mode),
 226                                  name, inum.inum, &dir_offset,
 227                                  STR_HASH_must_create);
 228         if (ret)
 229                 goto err;
 230
 231         inode_u->bi_dir         = dir.inum;
 232         inode_u->bi_dir_offset  = dir_offset;
 233
 234         ret =   bch2_inode_write(trans, &dir_iter, dir_u) ?:
 235                 bch2_inode_write(trans, &inode_iter, inode_u);
 236 err:
 237         bch2_trans_iter_exit(trans, &dir_iter);
 238         bch2_trans_iter_exit(trans, &inode_iter);
 239         return ret;
 240 }
 241
 242 int bch2_unlink_trans(struct btree_trans *trans,
 243                       subvol_inum dir,
 244                       struct bch_inode_unpacked *dir_u,
 245                       struct bch_inode_unpacked *inode_u,
 246                       const struct qstr *name,
 247                       bool deleting_subvol)
 248 {
 249         struct bch_fs *c = trans->c;
 250         struct btree_iter dir_iter = { NULL };
 251         struct btree_iter dirent_iter = { NULL };
 252         struct btree_iter inode_iter = { NULL };
 253         struct bch_hash_info dir_hash;
 254         subvol_inum inum;
 255         u64 now = bch2_current_time(c);
 256         struct bkey_s_c k;
 257         int ret;
 258
 259         ret = bch2_inode_peek(trans, &dir_iter, dir_u, dir, BTREE_ITER_intent);
 260         if (ret)
 261                 goto err;
 262
 263         dir_hash = bch2_hash_info_init(c, dir_u);
 264
 265         ret = bch2_dirent_lookup_trans(trans, &dirent_iter, dir, &dir_hash,
 266                                        name, &inum, BTREE_ITER_intent);
 267         if (ret)
 268                 goto err;
 269
 270         ret = bch2_inode_peek(trans, &inode_iter, inode_u, inum,
 271                               BTREE_ITER_intent);
 272         if (ret)
 273                 goto err;
 274
 275         if (!deleting_subvol && S_ISDIR(inode_u->bi_mode)) {
 276                 ret = bch2_empty_dir_trans(trans, inum);
 277                 if (ret)
 278                         goto err;
 279         }
 280
 281         if (deleting_subvol && !inode_u->bi_subvol) {
 282                 ret = -BCH_ERR_ENOENT_not_subvol;
 283                 goto err;
 284         }
 285
 286         if (inode_u->bi_subvol) {
 287                 /* Recursive subvolume destroy not allowed (yet?) */
 288                 ret = bch2_subvol_has_children(trans, inode_u->bi_subvol);
 289                 if (ret)
 290                         goto err;
 291         }
 292
 293         if (deleting_subvol || inode_u->bi_subvol) {
 294                 ret = bch2_subvolume_unlink(trans, inode_u->bi_subvol);
 295                 if (ret)
 296                         goto err;
 297
 298                 k = bch2_btree_iter_peek_slot(&dirent_iter);
 299                 ret = bkey_err(k);
 300                 if (ret)
 301                         goto err;
 302
 303                 /*
 304                  * If we're deleting a subvolume, we need to really delete the
 305                  * dirent, not just emit a whiteout in the current snapshot:
 306                  */
 307                 bch2_btree_iter_set_snapshot(&dirent_iter, k.k->p.snapshot);
 308                 ret = bch2_btree_iter_traverse(&dirent_iter);
 309                 if (ret)
 310                         goto err;
 311         } else {
 312                 bch2_inode_nlink_dec(trans, inode_u);
 313         }
 314
 315         if (inode_u->bi_dir             == dirent_iter.pos.inode &&
 316             inode_u->bi_dir_offset      == dirent_iter.pos.offset) {
 317                 inode_u->bi_dir         = 0;
 318                 inode_u->bi_dir_offset  = 0;
 319         }
 320
 321         dir_u->bi_mtime = dir_u->bi_ctime = inode_u->bi_ctime = now;
 322         dir_u->bi_nlink -= is_subdir_for_nlink(inode_u);
 323
 324         ret =   bch2_hash_delete_at(trans, bch2_dirent_hash_desc,
 325                                     &dir_hash, &dirent_iter,
 326                                     BTREE_UPDATE_internal_snapshot_node) ?:
 327                 bch2_inode_write(trans, &dir_iter, dir_u) ?:
 328                 bch2_inode_write(trans, &inode_iter, inode_u);
 329 err:
 330         bch2_trans_iter_exit(trans, &inode_iter);
 331         bch2_trans_iter_exit(trans, &dirent_iter);
 332         bch2_trans_iter_exit(trans, &dir_iter);
 333         return ret;
 334 }
 335
 336 bool bch2_reinherit_attrs(struct bch_inode_unpacked *dst_u,
 337                           struct bch_inode_unpacked *src_u)
 338 {
 339         u64 src, dst;
 340         unsigned id;
 341         bool ret = false;
 342
 343         for (id = 0; id < Inode_opt_nr; id++) {
 344                 /* Skip attributes that were explicitly set on this inode */
 345                 if (dst_u->bi_fields_set & (1 << id))
 346                         continue;
 347
 348                 src = bch2_inode_opt_get(src_u, id);
 349                 dst = bch2_inode_opt_get(dst_u, id);
 350
 351                 if (src == dst)
 352                         continue;
 353
 354                 bch2_inode_opt_set(dst_u, id, src);
 355                 ret = true;
 356         }
 357
 358         return ret;
 359 }
 360
 361 static int subvol_update_parent(struct btree_trans *trans, u32 subvol, u32 new_parent)
 362 {
 363         struct btree_iter iter;
 364         struct bkey_i_subvolume *s =
 365                 bch2_bkey_get_mut_typed(trans, &iter,
 366                         BTREE_ID_subvolumes, POS(0, subvol),
 367                         BTREE_ITER_cached, subvolume);
 368         int ret = PTR_ERR_OR_ZERO(s);
 369         if (ret)
 370                 return ret;
 371
 372         s->v.fs_path_parent = cpu_to_le32(new_parent);
 373         bch2_trans_iter_exit(trans, &iter);
 374         return 0;
 375 }
 376
 377 int bch2_rename_trans(struct btree_trans *trans,
 378                       subvol_inum src_dir, struct bch_inode_unpacked *src_dir_u,
 379                       subvol_inum dst_dir, struct bch_inode_unpacked *dst_dir_u,
 380                       struct bch_inode_unpacked *src_inode_u,
 381                       struct bch_inode_unpacked *dst_inode_u,
 382                       const struct qstr *src_name,
 383                       const struct qstr *dst_name,
 384                       enum bch_rename_mode mode)
 385 {
 386         struct bch_fs *c = trans->c;
 387         struct btree_iter src_dir_iter = { NULL };
 388         struct btree_iter dst_dir_iter = { NULL };
 389         struct btree_iter src_inode_iter = { NULL };
 390         struct btree_iter dst_inode_iter = { NULL };
 391         struct bch_hash_info src_hash, dst_hash;
 392         subvol_inum src_inum, dst_inum;
 393         u64 src_offset, dst_offset;
 394         u64 now = bch2_current_time(c);
 395         int ret;
 396
 397         ret = bch2_inode_peek(trans, &src_dir_iter, src_dir_u, src_dir,
 398                               BTREE_ITER_intent);
 399         if (ret)
 400                 goto err;
 401
 402         src_hash = bch2_hash_info_init(c, src_dir_u);
 403
 404         if (dst_dir.inum        != src_dir.inum ||
 405             dst_dir.subvol      != src_dir.subvol) {
 406                 ret = bch2_inode_peek(trans, &dst_dir_iter, dst_dir_u, dst_dir,
 407                                       BTREE_ITER_intent);
 408                 if (ret)
 409                         goto err;
 410
 411                 dst_hash = bch2_hash_info_init(c, dst_dir_u);
 412         } else {
 413                 dst_dir_u = src_dir_u;
 414                 dst_hash = src_hash;
 415         }
 416
 417         ret = bch2_dirent_rename(trans,
 418                                  src_dir, &src_hash,
 419                                  dst_dir, &dst_hash,
 420                                  src_name, &src_inum, &src_offset,
 421                                  dst_name, &dst_inum, &dst_offset,
 422                                  mode);
 423         if (ret)
 424                 goto err;
 425
 426         ret = bch2_inode_peek(trans, &src_inode_iter, src_inode_u, src_inum,
 427                               BTREE_ITER_intent);
 428         if (ret)
 429                 goto err;
 430
 431         if (dst_inum.inum) {
 432                 ret = bch2_inode_peek(trans, &dst_inode_iter, dst_inode_u, dst_inum,
 433                                       BTREE_ITER_intent);
 434                 if (ret)
 435                         goto err;
 436         }
 437
 438         if (src_inode_u->bi_subvol &&
 439             dst_dir.subvol != src_inode_u->bi_parent_subvol) {
 440                 ret = subvol_update_parent(trans, src_inode_u->bi_subvol, dst_dir.subvol);
 441                 if (ret)
 442                         goto err;
 443         }
 444
 445         if (mode == BCH_RENAME_EXCHANGE &&
 446             dst_inode_u->bi_subvol &&
 447             src_dir.subvol != dst_inode_u->bi_parent_subvol) {
 448                 ret = subvol_update_parent(trans, dst_inode_u->bi_subvol, src_dir.subvol);
 449                 if (ret)
 450                         goto err;
 451         }
 452
 453         /* Can't move across subvolumes, unless it's a subvolume root: */
 454         if (src_dir.subvol != dst_dir.subvol &&
 455             (!src_inode_u->bi_subvol ||
 456              (dst_inum.inum && !dst_inode_u->bi_subvol))) {
 457                 ret = -EXDEV;
 458                 goto err;
 459         }
 460
 461         if (src_inode_u->bi_parent_subvol)
 462                 src_inode_u->bi_parent_subvol = dst_dir.subvol;
 463
 464         if ((mode == BCH_RENAME_EXCHANGE) &&
 465             dst_inode_u->bi_parent_subvol)
 466                 dst_inode_u->bi_parent_subvol = src_dir.subvol;
 467
 468         src_inode_u->bi_dir             = dst_dir_u->bi_inum;
 469         src_inode_u->bi_dir_offset      = dst_offset;
 470
 471         if (mode == BCH_RENAME_EXCHANGE) {
 472                 dst_inode_u->bi_dir             = src_dir_u->bi_inum;
 473                 dst_inode_u->bi_dir_offset      = src_offset;
 474         }
 475
 476         if (mode == BCH_RENAME_OVERWRITE &&
 477             dst_inode_u->bi_dir         == dst_dir_u->bi_inum &&
 478             dst_inode_u->bi_dir_offset  == src_offset) {
 479                 dst_inode_u->bi_dir             = 0;
 480                 dst_inode_u->bi_dir_offset      = 0;
 481         }
 482
 483         if (mode == BCH_RENAME_OVERWRITE) {
 484                 if (S_ISDIR(src_inode_u->bi_mode) !=
 485                     S_ISDIR(dst_inode_u->bi_mode)) {
 486                         ret = -ENOTDIR;
 487                         goto err;
 488                 }
 489
 490                 if (S_ISDIR(dst_inode_u->bi_mode)) {
 491                         ret = bch2_empty_dir_trans(trans, dst_inum);
 492                         if (ret)
 493                                 goto err;
 494                 }
 495         }
 496
 497         if (bch2_reinherit_attrs(src_inode_u, dst_dir_u) &&
 498             S_ISDIR(src_inode_u->bi_mode)) {
 499                 ret = -EXDEV;
 500                 goto err;
 501         }
 502
 503         if (mode == BCH_RENAME_EXCHANGE &&
 504             bch2_reinherit_attrs(dst_inode_u, src_dir_u) &&
 505             S_ISDIR(dst_inode_u->bi_mode)) {
 506                 ret = -EXDEV;
 507                 goto err;
 508         }
 509
 510         if (is_subdir_for_nlink(src_inode_u)) {
 511                 src_dir_u->bi_nlink--;
 512                 dst_dir_u->bi_nlink++;
 513         }
 514
 515         if (dst_inum.inum && is_subdir_for_nlink(dst_inode_u)) {
 516                 dst_dir_u->bi_nlink--;
 517                 src_dir_u->bi_nlink += mode == BCH_RENAME_EXCHANGE;
 518         }
 519
 520         if (mode == BCH_RENAME_OVERWRITE)
 521                 bch2_inode_nlink_dec(trans, dst_inode_u);
 522
 523         src_dir_u->bi_mtime             = now;
 524         src_dir_u->bi_ctime             = now;
 525
 526         if (src_dir.inum != dst_dir.inum) {
 527                 dst_dir_u->bi_mtime     = now;
 528                 dst_dir_u->bi_ctime     = now;
 529         }
 530
 531         src_inode_u->bi_ctime           = now;
 532
 533         if (dst_inum.inum)
 534                 dst_inode_u->bi_ctime   = now;
 535
 536         ret =   bch2_inode_write(trans, &src_dir_iter, src_dir_u) ?:
 537                 (src_dir.inum != dst_dir.inum
 538                  ? bch2_inode_write(trans, &dst_dir_iter, dst_dir_u)
 539                  : 0) ?:
 540                 bch2_inode_write(trans, &src_inode_iter, src_inode_u) ?:
 541                 (dst_inum.inum
 542                  ? bch2_inode_write(trans, &dst_inode_iter, dst_inode_u)
 543                  : 0);
 544 err:
 545         bch2_trans_iter_exit(trans, &dst_inode_iter);
 546         bch2_trans_iter_exit(trans, &src_inode_iter);
 547         bch2_trans_iter_exit(trans, &dst_dir_iter);
 548         bch2_trans_iter_exit(trans, &src_dir_iter);
 549         return ret;
 550 }